Document

Report
RADIOMICS ON LUNG CANCER
Robert Gillies* (PI), Robert Gatenby* (Co-PI), Dmitry Goldgof#, Philippe Lambin##, Yoganand Balagurunathan*, Yuhua Gu*, Olya Grove*, Hua Wang*,Lawrence Hall#, Matthew Schabath*, Jhanelle Gray*,
Eduardo Moros*, Thomas Dilling*, Jongphil Kim*, Anders Berglund*, Steven Eschrich*, Gregory Bloom*, John Heine*, Andre Dekker##, Hugo Aerts##, Emmanuel Rios##
*MOFFITT
CANCER CENTER, #UNIVERSITY OF SOUTH FLORIDA,
Lung Screening
##MAASTRO
CLINIC (Netherlands)
Lung Cancer: Quantitative & Semantic Features
National Lung Screening Trial
Phantom Study
Lung$
Objective: We use NLST screening CT data and compute quantitative image features
and retrospectively predict nodule progression, especially for those cases developing
to malignant nodules.
Convexity Feature+$ :
+&:
Objective: The heterogeneity in sample collection due to clinical and imaging differences between samples can impact on translational radiomics
studies. It has to be noted that most clinical scans are obtained with different scan protocols. In order to quantify these changes and study the
impact on radiomic features, we used human tissue phantoms that was scanned on three different scanners (Phillips, Siemens, GE) at different tube
voltage (KVP). The Initial goal was to find features that were stable across the tube voltages and find the features that tracks density change.
Feature Report
Objective: One of the teams interest has been to develop new quantitative descriptors that show
prognostic value and are reproducible. We developed two imaging biomarkers from diagnostic CT
scans that capture the shape and texture changes within the tumor, using 2D stacks of CT scans.
Method: We quantitatively demonstrated in the subset of adenocarcinoma (N = 62) CT studies that the
developed biomarkers are capable of classifying patients based on CT scans alone into two prognostic
categories within statically significant bounds
Area=1201.6 mm 2
Perimeter=152.4 mm
Volume=2615 mm 3
Surf. area=1099.33 mm2
Surf. Area/volume=0.42 mm-1
Density, Necrosis
Mean = 19.33 HU
SD = 76.59 HU
Min = -249 HU
Max = 118 HU
Spiculations
Slope at margin =133.5±31.3 HU/mm
Low density inclusions
Rel. vol.= 0.21 mm 3
Number=110
Volume=4.89±8.35 mm 3
Longitudinal CT Image
Semantic Features +%:
(a)
Analysis
Above: Tissue Characterization Phantom, Gammex 467
Nodule Detection and Segmentation
Above (Left to Right): CT scanners (Phillips Brilliance, Siemens Sensation 64 and GE Discovery STE)
(Courtesy: Dr. Reeves, Cornell University)
Experimental Setup: The tissue phantoms were scanned on three scanners with three different voltages.
Due to scanner setting limitations tube currents were matched to close range between scanners.
Above: Radiomics on Lung Nodules
Challenges: The NLST data consists of over 26, 000 patients with CT image study.
Therefore selection of sub population is critical and the quantification process needs to
be efficient and automated. Area of focus include: Data management, Curation,
Segmenting the nodules (relatively small regions of interest).
Slice Thickness, Row Spacing,
Modality mm
mm
Column Spacing,
mm
kVp
GE
Phillips
3.27
3
0.97
0.844
0.977
0.844
Siemens
3
0.742
0.742
(c)
(a)
X-Ray Tube Current,
mA
100, 120
90, 120, 140
80, 100, 120,
140
(b)
Above Figures. Subdividing the tumor region (a) region of interest and (b) region between the markings is the boundary used
for the convexity measure (c) Kaplan-Meier plots for two groups formed by splitting the samples at the median value of
convexity score. Top plot are the patients with higher convexity values.
110, 80
203
402, 404, 414, 375
Above & Right: CT scanners setting and Phatoms scanned at different KVPs
Cohort Identification$. Based on the CT screening results (negative vs. positive) from
the baseline and two follow-up visits, we identified four distinct incident lung cancer
case cohorts (CCs). Two of the case cohorts developed lung cancer following one
(CC1) or two (CC2) prior positive screens that were not considered to indicate lung
cancer (CC1/2=192). CC1 patients had a positive screen at T0 then developed lung
cancer at T1; CC2 had positive screens at T0 and T1 then developed lung cancer at
T2. The other two case cohorts developed lung cancer following one (CC3) or two
(CC4) negative screens (CC3/4=125). CC3 had a negative screen at T0 then
developed lung cancer at T1; CC4 had negative screens at T0 and T1 then developed
lung cancer at T2. The prevalent cases (PCs) were diagnosed with lung cancer at
baseline (N=297).
T0
T1
N+ / C-
N+ / C-
N+ / C+
N+ / C+
N- / C-
Subcohort 1
(nodules w/ lung cancer):
Subcohort 2
(nodules w/o
lung cancer):
No. = 1,935
No. = 104
N+ / CN+ / C+
N- / CNo. = 14,563
N+ = Presence of a nodule
N- = No presence of a nodule
C+ = Lung cancer DX
C- = No lung cancer DX
N- / C-
N- / C-
N+ / C+
N+ / C+
+& Contributions: Drs. Budzevich & Moros
Figure Above. CV across different tube voltages (KVP) for features related to Pixel Intensity for Phillips, GE and Siemens scanner.
Objective: We evaluated reproducibility of 219 three-dimensional quantitative image features on the RIDER test/retest data set. A
subset of non redundant features was used to predict radiologist prognostic score. These reproducible set of features were then
tested for prognosis on a independent data set (59 Adenocarcinomas).
Above: Cohort selection in NLST Screening data.
CT screening results. Using the NLST definitions, positive screens are CT scans
that revealed any non-calcified nodule measuring ≥ 4 mm in any diameter and noncalcified nodule/mass(es) “suspicious for” lung cancer. Adenopathy or effusion
could be classified as a positive result. Abnormalities suggesting clinically
significant conditions other than lung cancer also were noted. Negative screens
were defined as CT scans with no significant abnormalities or significant or minor
abnormalities not suspicious for lung cancer.
$ (in preparation) Schabath et.al
Normal Lung Characterization %%:
Objective: We intend to study the quantitative image features characteristics across
the normal lung. This will also provide a base line set to study the repeatability of
the quantitative features.
Approach:
• The new designed lung field segmentation also includes bronchial tree
segmentation and vascular tree segmentation.
• After getting the pure Lung field segmentation, the next step is to divide the
lung field into 8 sub-volumes (1-level division), 32 sub-volumes (2-level
division) in order to extract texture features for each sub-volumes. (The level
of division depends on the need).
• This division process is trivial, (for each region (ex: left lung field), locate the
center of the region first and draw a Z-plane and Y-plane separately).
Concordance Correlation (ρc): It is defined as the ratio between the covered
range of a feature with respect to difference between repeats, and
standardized to cover [0 1].
E[(Y1  Y2 ) 2 ]
• Extract features for each sub-volume. 180 texture features for each region
includes: histogram, wavelet decomposition, laws feature, co-occurrence, runlength features. Shape features were not included.
Prediction of Radiologist Prognostic Score : The RIDER data was scored based
on fiver radiological semantic features, standardized to form two prognostic
groups : Better (over median) and Poor (below median). The quantitative
features were used to predict prognostic groups, of the repeatable features
conventional size & few texture based features showed predictive potential
(AUC > 0.7) using linear discriminant functions.
3
F1:LongDia
F2:ShortAxLongDia
F3:ShortAx
4
F6:Volume(cm3)
71.55
1
2
0.77
0.57
Specificity
AUC
0.82
0.88
0.86
0.89
0.79
0.88
0.87
0.91
Left: Different views of the lung
filed segmentation.
%% (in collaboration with) Chinese Academy of Sciences.
Above Figure : The Kaplan-Meier plots for two sample groups formed
by splitting the samples at the median values of their entropy value,
measured for (a) entire region of interest (b) the boundary (c) the core
of the tumors (d) between the core
Above. Kaplan-Meier survival plots for Pleural attachment & Lympho-adenopathy
features are shown to be significant in adenocarcinomas.
+% (in Review): Hua et.al, J. Thor. Onc.
Clinical Trial: Dasatinib
Objective: We hypothesize that Radiomics analysis would be able to identify image features that
would discriminate the cohort into responder and non-responders, retrospectively. These feature
may prospectively help us to better monitor the patients therapy and adjust the dosage levels.
• A variety of predictive classification models of survival were used including support vector
machines, Naive Bayes, rule based, and decision trees algorithms.
We achieve promising results of 77.5% accuracy with only 5 features using a Decision Tree
classifier
Classifier
Features
Decision Tree
Top 5 Relief-f
5
77.50%
65%
90%
0.712
Decision Tree
Top 10 Relief-f
10
70%
65%
75%
0.732
Rules
All
219
62.50%
65%
60%
0.729
Rules
Naive Bayes
All Top 5 Relief-f
All Top 10 Relieff
Manual &
Ensemble testretest (.85) Top 5
RF
SVM
All
Naive Bayes
AUC
0.93
Above. Runlength (GLN) feature discriminating
prognostic group. The top panel shows slice and 3D
of better prognostic tumor while bottom panel shows
poor prognostic group.
Samples: A cohort of 19 patients (12 Progressive Disease, 5
Stable Disease and 2 with Partial Response) has been in the
trial with Pre and Post-treatment CT scans, within 4 to 6
weeks of administering the drug.
#
Avg Accy
LQ Accy
UQ Accy
Approach: The CT images were segmented by a trained
radiologist with the help of CAD tools, Radiomics features
extracted. A best discriminating feature was identified based on
classifier analysis with various error estimations appropriate for
small samples. A combination of features would be appropriate
for discrimination.
Above: Waterfall plot of radiomic features that
tracks Stable (green) and progressive (red)
disease.
Database:
Objective: A mineable database with patient CT images, clinical parameters and quantitative
image features in a centralized location will accelerate discovery and collaboration among
researchers with-in and across the institutions.
AUC
5
75%
65%
85%
0.661
10
65%
55%
75%
0.52
5
60%
45%
75%
0.64
219
57.50%
60%
55%
0.575
Current Features:
•
An instance of ClearCanvas ImageServer has been installed as a research PACS
•
The Moffitt V2 data set with images, clinical data and image features has been loaded.
•
Additional CT scans has been sent to Clear Canvas for inclusion in validation and additional datasets
•
A easy-to-use web-based interface has been developed to manage the images in a project oriented
fashion.
•
Example of functionality includes, csv exports of all linked data, batch export of CT images to a work
directory for image analysis using Definiens or other software
•
Additional tools, such as XNAT, are continuously being evaluated for an improved ability to share the
data between institutions.
•
Major effort to curate and rename the NLST cohort is underway. This will ensure unique names to the
CT scans and will enable automatic retrial of the images through the Database.
0.79
Above Table: Summary of leave-one-out accuracy results containing the feature selection
method, number of features, average accuracy, lower quartile accuracy, upper quartile
accuracy, and area under the receiver operating curve. The top accuracy and AUC are
bold. The LQ represents Lower Quartile, UQ represents Upper quartile.
• Obtain natural range for each feature and reproducibility.
Independent Data: The repeatable and
informative features from the test/retest study
was used to test its prognostic potential in an
independent Non small cell lung cancer sample
set (59 Adenocarcinoma). The Runlength (GLN)
feature was found to be prognostic (P <= 0.046)
• Using the survival plots (KaplanMeier) we have found significance for
three semantic features: Pleural
attachment (P < 0.01),
Lymphadenopathy (P =0.03) and AirBronchogram (p=0.02).
• Using Kaplan-Meier survival analysis, we
evaluated the performance of entropy-based
imaging biomarker.
•
Redundancy Reduction (R2 ): Dependency between the features is quantified by
Coefficient of determination. To find an informative set, the featured are grouped at
different levels of R2 and replaced by a feature with highest DR.
Accurac Sensitivity
y
(percen
t)
81.03
0.8
80.17
0.75
77.59
• we not only computed the statistics based on
the entropy filtering across the entire tumor
region for all the image slices containing the
tumor, but also kept the record for the values
in the predefined core and edge sub-regions.
• Testing was done on 40 NSCLC adenocarcinoma cases. Relief-F was used to identify useful
image features from an initial set of 219 3D features.
 12   22  1  2 2
C5: Grayscale: Runlength & CoOccurrence
Feature
Accur Sensitivity Specifici
ty
acy
(perce
nt)
1 F48:AvgRunL(
70.69 0.47
0.96
GLN)
2 F51:AvgRunL(
73.28 0.67
0.80
LRE)
6
Results:
Figure Above. Process flow in the Test/Retest study
C1: Tumor Size
Feature
e) Associated findings (4):
Nodules in primary tumor lobe, Nodules in non-tumor lobes, Lymphadenopathy, Vascular
invasion.
• We used 109 non small cell Lung
cancer samples with
Adenocarcinomas to score the
semantic features.
Objective: We use CT-scan images to develop predictive models for lung cancer patient survival
time. Patients afflicted with Non-Small Cell Lung Cancer (NSCLC) were studied towards the
development of providing a novel approach to non-invasive tumor analysis.
Dynamic Range (DR): It is defined as the ratio between the covered
range of a feature with respect to difference between repeats, in the
given sample set. Standardized to [0 1].
c  1 
• we believe that by evaluating the difference
between the core and the boundary subregions, we made the feature less sensitive to
the inconsistencies in image formation.
Survival Prediction #^
Representative Features: The repeatable features that show high
concordance between test/retest experiments, large dynamic range and
non-dependent are the needed characteristics for a reliable biomarker, we
referred to such a subset as representative features.
Features: We have developed 25 semantic features which is broadly categorized into five groups:
a) Tumor (16):
Space localization (4): Location, Distribution, Fissure/Pleural attachment.
Size (1)
Shape (4): Sphericity, Lobulation, Concavity, Irregularity.
b) Margin (2):
Border definition, Spiculation
c) Density (5):
Texture, Air space, Air bronchogram, Calcification, Enhancement heterogenity
d) Surrounding tissue (5):
Pleural retraction, Vascular convergence, Thickened adjacent bronchovascular bundle,
Emphysema/Fibrosis periphery.
Experimental Setup & Results:
+$ (in Review): Grove et.al, J. Thor. Onc.
Quantitative Features: On the segmented tumor region of interest (ROI)
about 219 three dimensional features were extracted. The features are
broadly divided into seven category that describe the shape, size, location
and texture of the tumor.
Subcohort 3 (“normals”):
N- / C-
Location-specific entropy-based markers:
Feature Repeatability+# & Prognosis +# # :
T2
N+ / CN+ / C+
N- / C-
Results: About 31 features had coefficient of variation (CV) less than 10% , 10 feature in the
range of 10 to 20%CV, while about 178 features have CV over 20% for different KVP’s and
scanners.
Objective: The conventionally observed radialogical characteristics in a CT Lung images are referred
to as semantic features. We have converged on 25 most relevant features to describe the lung lesion
and developed a semantic scoring scheme on a five point scale.
+# Balagurunathan et.al, J. Dig. Imaging, 2013(in review)
+## Balagurunathan et.al, J. Tran. Oncology, 2014 (Accepted, in press)
#^(submitted):
S. Hawkins, et.al,USF
Acknowledgement:
This work is supported by NIH U01 CA143062 and Florida State’s James & Esther King Grant #2KT01

похожие документы